Identifying and assessing windbreaks in Ford County, Kansas using object-based image analysis

(1)

IDENTIFYING AND ASSESSING WINDBREAKS IN FORD COUNTY, KANSAS USING OBJECT-BASED IMAGE ANALYSIS

by

MIKE W. DULIN

B.S., KANSAS STATE UNIVERSITY, 2007

A THESIS

submitted in partial fulfillment of the requirements for the degree MASTER OF ARTS

Department of Geography College of Arts and Sciences

KANSAS STATE UNIVERSITY Manhattan, Kansas

2009

Approved by: Major Professor Dr. J.M. Shawn Hutchinson

(2)

Abstract

Windbreaks are a valuable resource in conserving soils and providing crop protection in western Kansas and other Great Plains states. Currently, Kansas has neither an up-to-date inventory of windbreak locations nor an assessment of their condition. The objective of this study is to develop remote sensing and geographic information system methods that rapidly identify and assess the condition of windbreaks in Ford County, Kansas. Ford County serves as a pilot study area for method development with the intent of transferring those methods to other counties/regions in Kansas and the Great Plains. A remote sensing technique known as object-based classification was used to classify windbreaks using color aerial photography acquired through the 2008 National Agricultural Imagery Program. Object-based classification works by segmenting imagery where areas with similar spectral, shape, and textural properties are grouped into vectors (i.e., objects) that are later used as the basis for image classification. Using this technique, 355 windbreaks, totaling nearly 1,012 acres (410 hectares), were identified in Ford County. When compared to a spatial dataset of confirmed windbreak locations generated via a heads-up digitizing process, the location of windbreaks identified using object-based classification results agreed approximately 81% of the time. Mean textural and spectral values were then combined and used to place identified windbreaks into three condition categories (good, fair, and poor) using a manual classification approach. Analysis showed the area of windbreaks in good condition to be 170 hectares, with the remaining 171 hectares of windbreaks falling in the fair or poor classes. Methods detailed in this study proved successful at rapidly identifying windbreak location and for providing useful condition class results for windbreak renovation and restoration planning.

(3)

List of Figures

Figure 1. Field windbreak located 20 miles east of Dodge City, Kansas. ... 1

Figure 2. Field windbreak located approximately 25 miles east of Dodge City Kansas. ... 2

Figure 3. Classification diagram used by Read (1958) to assess the condition of windbreaks. .. 11

Figure 4. A visual example of the segmentation process... 16

Figure 5. Differences between the main and first-order lines of an object... 19

Figure 6. Ford County study area with Kansas Forest Service districts highlighted... 25

Figure 7. Landcover/Landuse classes in Ford County... 27

Figure 8. Graph of population change in Ford County, 1870-2009... 29

Figure 9. Five year precipitation averages for Ford County, 1895-2006... 30

Figure 10. Five year mean temperatures for Ford County, 1895-2006. ... 31

Figure 11. Workflow sketch of the methods used in this research. ... 33

Figure 12. Comparison of the Ford County 2008 NAIP image at 1 and 6 meter pixel size... 36

Figure 13. Location of Ford County test site. ... 37

Figure 14. Screenshot of the ENVI Zoom 4.5 Feature Extraction Module and viewing portal. . 38

Figure 15. Screenshot of the band thresholding operation used to eliminate areas not required for windbreak classification... 41

Figure 16. Classified landuse/landcover map of the features generated in Ford County using object-based classification. ... 51

Figure 17. Highly erodible and impaired soils in Ford County, Kansas. ... 60

Figure 18. Segmented image of vegetative features in Ford County (1:235,000 scale)... 62

Figure 19. Portion of the segmented image of vegetative features for Ford County (1:60,000 scale). ... 63

Figure 20. Images displaying accurate windbreak border assignment and an inaccurate border assignment... 66

Figure 21. Example of a fragmented windbreak classified by NRCS as a single unit, but as two distinct units by object-based classification... 68

Figure 22. Box-and-whisker plot illustrating the overlap in NDVI values across windbreak condition classes for field-surveyed windbreaks. ... 72

(6)

Figure 23. Box-and-whisker plot illustrating the variation in textural values across windbreak condition classes for field-surveyed windbreaks. ... 76 Figure 24. Example of windbreak classified during a field survey as being in good condition but

in poor condition when using the standard deviation classification technique... 79 Figure 25. Example of a fair condition windbreak containing several areas that need renovation

(7)

List of Tables

Table 1. Common windbreak tree species in Ford County, Kansas. ... 9

Table 2. Total area of each landuse/landcover class in Ford County (Source: 2001 NLCD)... 28

Table 3. Historic population change in Ford County, Kansas. ... 28

Table 4. Annual precipitation for Ford County, 2000-2008. ... 30

Table 5. Custom landuse/landcover classes developed based on expert knowledge of the features remaining in the Ford County image after band thresholding. ... 44

Table 6. Discrete Capability Index overall interval types... 45

Table 7. Example of a populated error matrix with generic variables... 48

Table 8. Parameter settings used to classify the Ford County 2008 NAIP image. ... 49

Table 9. Final landuse/landcover categories used in the Ford County classification. ... 50

Table 10. Number of training sites selected for each landuse/landcover class in Ford County. . 50

Table 11. Software processing times required for each step in the classification of the Ford County NAIP image using object-based methods. ... 52

Table 12. Windbreak condition class evaluation guide. ... 58

Table 13. Windbreak attributes used to place windbreaks into condition classes. ... 59

Table 14. Classification error matrix for the Ford County test site. ... 64

Table 15. Error matrix compiled from a random distribution of sample points for the Non-Windbreak (NWB) and Non-Windbreak classes. ... 65

Table 16. Comparison of windbreak acres for the NRCS and object-based classification products... 67

Table 17. Error matrix for windbreak condition assessment based on average green reflectance and a quantile classification scheme. ... 69

Table 18. Error matrix for windbreak condition assessment based on average green reflectance and a natural breaks classification scheme. ... 70

Table 19. Error matrix for windbreak condition assessment based on average green reflectance and a standard deviation (SD) classification scheme... 71

(8)

Table 20. Error matrix for windbreak condition assessment based on average NDVI values and a quantile classification scheme... 73 Table 21. Error Matrix for windbreak condition assessment based on average NDVI values and a

natural breaks classification scheme... 73 Table 22. Error matrix for windbreak condition assessment based on average NDVI values and a

standard deviation (SD) classification scheme. ... 73 Table 23. Error matrix for windbreak condition assessment based on mean texture values and a

standard deviation (SD) classification scheme. ... 76 Table 24. Error matrix for windbreak condition assessment based on average condition index

values and a standard deviation (SD) classification scheme... 77 Table 25. Windbreak area by condition class. ... 78 Table 26. Total area of Ford County windbreaks located on highly erodible and impaired soils by condition class... 79

(9)

List of Equations

Equation 1. Full Lambda-Schedule algorithm for segment merging... 39

Equation 2. Method used to distinguish between classes by DCI interval type. ... 45

Equation 3. The Fitzpatrick-Lins equation. ... 47

Equation 4. The Cohen Kappa statistic... 49

Equation 5. Normalized Difference Vegetation Index (NDVI)... 55

Equation 6. Formula to calculate area (m2) for each windbreak condition class... 57

Equation 7. Formula to convert area to acres from square meters... 57

Equation 8. Formula to convert acres to hectares... 57

(10)

Acknowledgements

I would like to thank Dr. J.M. Shawn Hutchinson (Department of Geography, Kansas State University) for a tremendous amount of input and advice pertaining to this manuscript. I would also, like to thank Dr. Hutchinson for giving me the opportunity to work in the Geographic Information Systems and Spatial Analysis Laboratory (GISSAL) in the Department of Geography at Kansas State University. The valuable work experience I received in GISSAL made this project possible and finally got me a job working for Uncle Sam at the Army Corps of Engineers.

Dr. John Harrington Jr. and Dr. Douglas Goodin, who served as my two additional committee members, deserve special thanks for all of their advice and guidance which helped lead to the completion of this manuscript.

Funding for this research was provided from the Kansas Forest Service (KFS) through a USDA State and Private Forestry FY 2008 Western Competitive Grant. Thanks to Robert Atchison (KFS) this grant was successfully funded for years 2008-2010.

Finally, I would like to thank my fiancée Danielle White for her understanding, patience, and love as I had to dedicate more time to my work than her. Thanks for sticking with me Danielle!

(11)

Dedication

This thesis is dedicated in loving memory to the late Pamela Dulin (Dec. 31, 1949 – July 22, 2005). You were a wonderful mother and are greatly missed. It was through your guidance and dedication as a mother that my college education was made possible.

(12)

CHAPTER 1 - Introduction

Windbreaks, also known as shelterbelts, are a valuable resource in conserving soil and providing crop protection in western Kansas, as well as many other Great Plains states (Brandle et al., 2004). Many of these windbreaks were planted to reduce wind erosion during the Dust Bowl era of the 1930’s (Read 1958). The Society of American Foresters (SAF) defines a windbreak as “a strip of trees or shrubs maintained mainly to alter wind flow and microclimates in the sheltered zone, usually farm buildings.” SAF also defines a shelterbelt as a “strip of trees or shrubs maintained mainly to alter wind flow and microclimates in the sheltered zone, usually agricultural fields” (R. Atchison 2008, pers. comm.). Brandle et al., (2004) provide a much simpler definition “Windbreaks or shelterbelts are barriers used to reduce wind speed.” In this study, the Brandle et al., (2004) definition will be used as both farmstead and field windbreaks are objects of interest. Figures 1 and 2 are photos of multi-row windbreaks commonly found throughout the Ford County study area.

(13)

Figure 2. Field windbreak located approximately 25 miles east of Dodge City Kansas.

Currently, Kansas has neither an up-to-date inventory of windbreak locations nor an assessment of their condition. In this study, methods are applied to identify the location and assess the condition of windbreaks in Ford County, Kansas using remote sensing (RS) and geographic information system (GIS) techniques. The most recent study in Kansas was completed in 1992, when the U.S. Department of Agricultural (USDA) Natural Resource Conservation Service (NRCS) determined there were approximately 78,000 windbreaks in the state (R. Atchison 2008, pers. comm.). Most natural resource professionals agree that field windbreaks are a declining resource in Kansas and that few new ones are being established, yet there is little good science available to efficiently document windbreak location, size, or condition. Still, the 1997 NRCS Natural Resource Inventory (NRI) suggests that wind continues

(14)

to erode 1.8 million acres (728,434 ha) of cropland in Kansas at rates that exceed tolerable limits (around 1.3 tons/acre/year) (R. Atchison 2008, pers. comm.).

Sorenson and Marotz (1977) expressed concerns that windbreaks in Kansas were beginning to be removed over 3 decades ago. Many of the windbreaks planted in western Kansas were planted during the Dust Bowl era whenCongress passed the Prarie States Foresty Act (PSFA) (Croker 1991). The Act called for the planting of millions of trees and tens of thousands of shelterbelts in an attempt to prevent eolian erosion and to create jobs for a destitue Great Plains economy (Read 1958). Bates (1924) pointed out that, before the Dust Bowl, the lack of windbreaks in the Great Plains was a “severe handicap” to agricultural land. It is now timely to build upon previous work of the NRCS and additional shelterbelt research projects to develop methods for the rapid identification of windbreak location and an assessment their condition..

Purpose and Objectives

The main purpose of this research is to work in conjunction with the Kansas Forest Service (KFS) to develop GIS and RS methods to identify the location, size, and condition of windbreaks in Ford County, Kansas. In addition, it is the purpose of this research to determine how well automated classification schemes match with ground truth data of windbreak condition. Once a satisfactory method of classification has been obtained, the same or similar methods can be used in future research to locate and assess windbreaks in other western Kansas counties and across the Great Plains.

(15)

This project has three primary goals:

1) To rapidly classify windbreaks using object-based classification on County Composite Mosaic (CCM) aerial imagery,

2) To develop a secondary classification to assess windbreak condition (good, fair, poor) and determine the number of acres/hectares that exist in each class, and

3) To identify windbreaks located on highly erodible and impaired soils.

To achieve these goals a remote sensing technique known as object-based classification was used to classify windbreaks from 2008 National Agriculture Imagery Program (NAIP) imagery. Attributes from four spectral bands (blue, green, red, and near-infrared) present in the NAIP imagery were used in the classification process. Object-based classification takes into account size, shape, and context as well as spectral information of features identified for classification (Baatz et al., 2004). These non-spectral classification criteria are crucial for accurate classification of windbreaks for two reasons. First, windbreaks are usually linear strips of tree plantings. A riparian area could easily share similar spectral reflectance characteristics as a windbreak making it difficult to distinguish between the two features without considering some shape criteria in the classification. Second, some object-based classification software packages allow for the isolation of features of interest. This option means that based on certain shape and spectral parameter settings, one can eliminate features in the image that are not of interest before beginning the classification process, resulting in more efficient classification and image processing times.

Traditionally, foresters determine windbreak condition using measures of tree density and/or porosity. In addition, windbreak condition can also be based on how well it is functioning as a wind restraint. A dense windbreak is capable of blocking more wind from the sheltered

(16)

zone while a sparsely-planted windbreak provides less wind resistance over the same area. Read (1958) developed condition classification criteria that are still widely accepted and used (R. Atchison pers. comm. 2008). Read’s classification approach requires field crews to visit each windbreak and look through it horizontally to determine density. In this study, windbreak condition is assessed using a vertical view provided by aerial photographs. If assessment results generated by analysis of aerial photography compare well with ground-based assessments, many hours of field surveying could be replaced with digital image processing for condition assessments.

The mean spectral reflectance brightness values (BV) of windbreak features in the green band (band 2) of windbreaks were used to determine windbreak condition. The average BV of a windbreak should correlate well with vegetation density and tree cover within a given stand. Dense windbreaks appear as dark linear features with few gaps where bare soil and grass is visible through the tree canopy. Such windbreaks should be in relatively good condition and have a lower average BV as compared to a less dense windbreak with multiple or large gaps in the stand.

Normalized Difference Vegetation Index (NDVI) was also calculated from the NAIP image to perform a second condition assessment and then be compared with results from the BV-based assessment. As a measure of greenness, average NDVI values for a windbreak should be a good measure of vegetation abundance and health. A third, and final, condition assessment was performed using a condition index based on both mean textural values and BV’s using a linear scaling model suggested by Booysen (2002).

In addition to knowing the location and associated condition of windbreaks, Kansas foresters are also interested in understanding more about the soils on which they are found. To

(17)

achieve the third goal of this project, a geospatial data layer of highly erodible and impaired soils was obtained from Hutchinson et al., (2008). This layer was composed of agricultural crop land and soils with a wind erodibility index (WEI) of 87 or greater. Identifying windbreaks in these vulnerable areas can assist foresters in prioritizing windbreak renovation projects and new areas for future plantings.

Applied and Academic Contribution

Results from this research will assist foresters in Kansas as they plan for future windbreak renovations and identify sites for new plantings. In addition, this research may serve as a catalyst for future research that places a monetary value on the ecosystem services provided by windbreaks. Within the academic realm of geography this research helps answer one of the most common questions in the discipline, ‘where?’(Pattison 1963; Golledge 2002; Cutter et al., 2004). Cutter et al., (2004) identify ten “Big Questions in Geography” that are meant to work as guides in helping bring geographic research to the public in meaningful and useful ways. Research from this project addresses two of these questions:

1) “How has the earth been transformed by human action?” and 2) “What role will virtual systems play in learning about the world?”

Prior to the 1930’s few, if any, windbreaks existed in Kansas or the Great Plains (Croker 1991). Identifying windbreaks will contribute to the understanding of how humans have modified the landscape in the interest of soil loss prevention. Additionally, use of RS and GIS methods is likely to promote better understanding of ‘what’ we can identify ‘where’ on the landscape. Understanding the capabilities and limits of virtual systems is a fundamental element

(18)

when determining the scenarios for which these systems can be effectively used to extracting valid information from real-world imagery.

(19)

CHAPTER 2 - Literature Review

Windbreaks

Windbreaks provide a number of environmental benefits for semi-arid regions throughout the world. Windbreaks provide protection from wind for cattle, crops, and soil as well as homes and other structures (Kort and Stefner 2007; Brandle et al., 2004). Stoeckeler and Williams (1949) reported many farmers throughout the Great Plains encountered a drastic reduction in winter fuel expenses after planting windbreaks to protect their homes.

Windbreaks first appeared on the landscape in Scotland during the mid-1400’s after Scottish Parliament urged planting to assist in soil loss prevention on agricultural lands (Brandle et al., 2004; Droze 1977). The first major planting of windbreaks in the United States occurred during the 1930’s. Due to Dust Bowl conditions, U.S Congress authorized the Prairie States Forestry Project (PSFP) to assist with planting of shelterbelts to minimize wind erosion and decrease the number and intensity of dust storms (Brandle et al., 2004; Droze 1977; Read 1958). Many of the dust storms common to Kansas were the result of severe drought, mismanaged agricultural land, and “suit case farming” practiced by transient farmers (Saloutos 1969 pg. 1). The PSFP provided jobs for an economically distressed population and resulted in the planting of over 200 million trees and shrubs totaling 18,600 miles in length (Read 1958). Between 1935 and 1942, the PSFP was successful in planting shelterbelts on 30,000 farms stretching from the Canadian border of North Dakota south to the Texas Panhandle (Read 1958). Windbreaks planted in western Kansas consisted mainly of the species listed in Table 1.

Controversy arises, however, when dealing with the total number of trees actually planted by the PSFP. Croker (1991) stated that only 145 million trees were planted by the PSFP between

(20)

1935 and 1942, which is not consistent with the numbers given by Read (1958). Though the numbers do not match, there is little argument that the PSFP was “One of the greatest projects ever attempted by man to improve the environment of our planet” (Croker 1991 pg. 4).

Table 1. Common windbreak tree species in Ford County, Kansas. Scientific Name Common Name Prunus angustifolia Sand Hill plum

Prunus americana American Plum

Prunus virginiana Choke Cherry

Fraxinus pennsylvanica Green Ash

Populus deltoides Eastern Cottonwood

Ulmus pumila Siberian Elm

Celtis occidentalis Hackberry

Gleditsia triacanthos Honey Locust

Quercus macrocarpa Bur Oak

Gymnocladus dioicus Kentucky Coffee Tree Juniperus scopulorum Rocky Mountain Juniper Juniperus virginiana Eastern Red Cedar

Pinus ponderosa Ponderosa Pine

Pinus sylvestris Scotch Pine

Pinus nigra Austrian Pine

Previous Windbreak Assessments

In 1938, the first windbreak survey in the United States was conducted to assess the survival rate of windbreaks as a whole and not just the individual health of trees planted during the PSFP. Results showed that 61% of all species planted survived (Read 1958). Factors cited for the failure of the remaining 39% of plantings included poor agricultural practices, insects, rodents, and improper planting strategies (Read 1958). While this survey reported a fairly high survival rate for plantings, it did not take into consideration the condition of surviving windbreaks.

In forestry terms, windbreak condition is determined based on how well it is functioning rather than solely on the health of individual trees. Function ratings are based on the number of

(21)

gaps (i.e., porosity) in a windbreak (Cornelis and Gabriels 2005). Porosity can be measured as a ratio between the open area and the total area of a windbreak (Cornelis and Gabriels 2005; Jensen 1954).

A second survey was conducted in 1944 to assess PSFP-planted windbreaks. By this time, the plantings ranged in age from 4 to 7 years (Read 1958). Samples included over 1,000 windbreaks spanning from North Dakota to Texas with 78% of them being rated in good condition or higher (Read 1958). Read (1958) explains that the criteria for classifying a windbreak as “good” was based only on “survival and the potential of producing a barrier”.

In 1954, Read (1958) conducted another survey that re-examined many of the same windbreaks (938 or 1,079) sampled ten years earlier. His research took into consideration survival, height, diameter at breast height (DBH), vigor, crown spread, and continuity of trees for each species. In addition, Read (1958) also developed four classes (good, fair, poor, or destroyed) into which windbreaks could be placed based on their “effectiveness”, which is now considered to be “function” (Figure 3). Good windbreaks in the Read (1958) study exhibited moderate but continuous density throughout the stand. Windbreaks in the poor class generally had low density with sparsely separated or clumped trees. Read (1958) reported that 42% of the windbreaks surveyed were in good or excellent condition, 31% were classified as fair, 19% classified as poor, and the remaining 8% had been removed (Ticknor 1989).

More recently, the 1992 NRCS Natural Resource Inventory (NRI), the last assessment conducted in Kansas, identified 78,000 windbreaks in the state. These windbreaks covered a total of 114,000 acres and extended a collective length of 20,000 miles. Of those windbreaks surveyed, 13% were found to be in excellent condition, 38% good, 34 % fair and 15% poor (United States Department of Agriculture 1994). In producing the 1992 NRI, NRCS adopted

(22)

many of the same criteria for classifying windbreaks as those used by Read (1958). Read’s classification scheme remains the most widely accepted and used method of rating windbreaks in the field (R. Atchison 2008 pers. comm.).

Figure 3. Classification diagram used by Read (1958) to assess the condition of windbreaks.

(23)

Recently, Wiseman et al., (2007) used object-based image analysis using very-high resolution (VHR) imagery to identify and assess windbreaks. Segmentation methods were used on VHR images to create a spatial polygon file which was then used to extract windbreaks via Structured Query Language (SQL) queries with in a GIS. A total of 27 sample windbreaks were selected throughout a study area located approximately 150 kilometers east of Winnipeg, Manitoba, Canada. Of these 27 windbreaks, 26 were classified correctly when compared to ground truth data. Wiseman et al., (2007) also used spectral properties from the red, green, and blue bands to determine windbreak density and shape characteristics along with expert knowledge to extract windbreak vectors from their segmented data.

Remote Sensing

Remote sensing is commonly used for the identification, extraction, and classification of landuse/landcover types (Koch et al., 2007). Biophysical remote sensing techniques have also been proven useful in monitoring vegetation biomass, soil moisture, surface temperature, and surface texture (Jensen 1983). At the simplest level, remote sensing can be thought of as the process of extracting data from real world imagery (Quattrochi et al., 1989). Geographic remote sensing requires that users understand the nature of the imagery they are working with (i.e., knowing what landcover features they are looking at) and the drivers or physical processes that determine why particular features or cover types are located where they are (Quattrochi et al., 1989).

Remote sensing methods take advantage of advanced sensors to capture images of particular features or geographic areas. The process of gathering remotely-sensed imagery has been around for over 150 years (Jensen 2007). In 1858, the first aerial image was captured by the Frenchman Nadar from a tethered balloon (Jenson 2007). Modern sensors are capable of

(24)

capturing imagery at multiple angles, various spatial resolutions, and using most areas of the electromagnetic spectrum. These advanced sensors can be found on satellites orbiting the earth, onboard terrestrial aircraft, and even contained within portable units that users can take to the field to capture detailed information about the earth’s surface (Rostoker et al., 1995; Moran et al., 1997; Diner et al., 1999; Landgrebe 2003; Jensen 2007).

Imagery collected through remote sensing is not only used for visual interpretation, but also to extract relevant thematic information through classification and to analyze various biophysical properties of vegetative land cover (Jensen 2005). Extracting thematic information from imagery is typically accomplished through ‘supervised’ or ‘unsupervised’ classification approaches (Jensen 2005; Richards and Xiuping 2005; Schowengerdt 2007).

In a supervised classification, known ground truth points are selected as training sites for a predefined number of thematic classes. These training sites are then used in classification algorithms to group like clusters of homogeneous pixels into their respective class (Jensen 2005; Richards and Xiuping 2005). In unsupervised classifications, algorithms are used to group like pixels into categories with little, or no, prior knowledge of the thematic types present in the imaged area (Duda et al., 2001; Jensen 2005). Once the unsupervised classification scheme groups like pixels, the user then labels the groups or classes according to their corresponding information class as determined from expert knowledge or ground truth data (Famiglietti et al., 1999; Jensen 2005).

Both supervised and unsupervised classification methods have traditionally been accomplished on a per-pixel basis. Per pixel classifications takes into account only the spectral value of a single pixel, which limits its capability to identify ‘features’ and process VHR data (Jensen 2005). More recently, object-based classification methods have been gaining in

(25)

popularity. Object-based classifications group homogeneous pixels through a segmentation process and convert them to multi-pixel shapes which later become the basis for classification (Blashke et al., 2004; Jensen 2005). Objects created during segmentation provide not only spatial information, but also have spectral and textural properties of the objects associated with them (Baatz et al., 2004). For example, each segmented object will have an associated mean pixel value for each electromagnetic waveband as an attribute which can be used later during classification.

Object-based Classification Concepts

Object-based classification is not a new idea. White the approach has been known since the 1970’s, a lack of computing power has prevented its widespread use until very recently (Rutherford and Rapoza 2008). Traditional pixel-based classification only takes into account position, size, and value of individual pixels in remotely-sensed imagery (Jenson 2005). Object-based classification takes into account not only the position, size, and spectral characteristics of individual objects, but also shape and context as a forth category to delineate between individual landscape features, or objects (Blaschke et al., 2004). Essentially, object-based classification allows a classification scheme to be based on the shape of objects or features rather than simply the spectral reflectance of single pixels. The main purpose of object oriented-image analysis is to extract or identify “real world objects” that are “proper in shape and proper in classification” (Baatz et al., 2004).

Object-based classification depends on a critical process called segmentation. Segmentation groups similar pixels together to create object vectors. Conceptually, segmentation methods are used to divide an image into homogeneous objects or regions for extraction and classification (Koch et al., 2007). However, a “general segmentation method,

(26)

which performs well in many contexts, does not exist” (Kermad and Chehdi 2002 pg. 542). This means that a general scale parameter setting, which determines the size of objects to be created, does not exist to segment all desired objects all the time in remotely-sensed images. In the context of this study, if the scale-based segmentation parameter is set too high, the image may be segmented into very large polygons that encompass much more area than the features of interest (e.g., windbreaks). Alternatively, if the scale parameter is set too low, then the image could be segmented the component pieces of a feature of interest (e.g., individual trees in a windbreak) and fail to capture the entire feature as one object. Figure 4 is provided to facilitate visualization of a segmented aerial image.

In order to overcome the issues surrounding scale-based segmentation parameter settings, many parameter values need to be tested to determine the optimal segmentation size(s) based on the type of object(s) a user wants to classify (Rutherford and Rapoza 2008). This approach introduces some subjectivity into the segmentation process but little has been done to generalize segmentation parameters in the interest of isolating any given set features.

There are three basic and accepted approaches to performing image segmentation: Pixel, edge, and region (Blaschke et al., 2004). Pixel-based approaches assign a label to continuous patches of pixel cells. Edge-based methods attempt to identify edges between regions and assign labels to boundaries between regions where a pixel value change occurs. According to Robinson et al., (2002), the edge-based approach works fastest because it incorporates only the scale parameter into the segmentation process. Region and pixel-based segmentation methods incorporate compactness (ratio of pixels in the perimeter length) and smoothness (jagged edges vs. smooth edges) into the segmentation algorithm which slows down processing time (Kermad and Chehdi 2002; Jensen 2005).

(27)

Figure 4. A visual example of the segmentation process.

Original Segmented

Grey Scale Objects

Region-based techniques are subdivided into three categories: Region growing, region merging, and region splitting (Baatz et al., 2004). Seed points are used to start this region-based process. Next, neighboring pixels are joined to the original seed point until a specific size threshold is met. After this, the process starts over with a new set of seed points (Blaschke et al., 2004). Kermad and Chehdi (2002) suggest that an integrated segmentation approach will yield the best results, especially when edge- and region-based approaches are combined. Often, a

(28)

quite successful. This approach, referred to as the fractal net evolution approach (FNEA), is inherent in the eCognition (Definiens, Muchen Germany) digital image segmentation software (Huang and Zhang 2008).

The ENVI Zoom 4.5 Feature Extraction Module (ITT Visual Information Solutions, Boulder, Colorado), which was used in this research, uses an edge-based algorithm developed by Robinson et al., (2002) to segment imagery. However, little research has been published using this software due to its recent availability (ITT Visual Information Solutions 2008). The edge-based algorithm is intended to work very fast because it requires only one parameter input (scale level) (ITT Visual Information Solutions 2008). Robinson et al., (2002) algorithm was developed to detect edge features then merge neighboring regions based on similar measures of spectral values as scale parameter increases.

Expert knowledge, or knowledge-based image interpretation, is a strategy in which a user determines the proper segmentation and classification parameters based on his/her knowledge of the objects they wish to classify (Rutherford and Rapoza 2008; Benz et al., 2004). For example, users can look at an image and determine if the segmentation parameters did or did not adequately capture the objects of interest. Benz et al., 2004 (pg. 241) list four key points to employ the best knowledge-based image interpretation techniques:

1. Understanding sensor characteristics,

2. Understanding the appropriate scale of analysis,

3. Identification of typical context and hierarchical dependencies,

4. Consideration of inherent uncertainties of the whole information extraction system, starting with the sensor, up to fuzzy concepts for the requested information.

(29)

Complementary to expert knowledge, is the ability of object-based classification to account for the compactness, smoothness, and linearity of various objects (Tian and Chen 2007; Baatz et al., 2004). Linearity can be represented by three different object feature measures: (1) Length/width of object, (2) length/width of main line, and (3) length/width of main line plus long branches of first order” (Tian and Chen 2007). Tian and Chen (2007) suggest that using length/width of main line is far superior to length/width of object because it can capture the true linearity of objects. Figure 5 illustrates a generic main line and first order lines of a generic object. Baatz et al., (2004) mentions, however, that relying too much on shape information may reduce the quality of segmentation and advises users to emphasize spectral information while incorporating shape information only when necessary.

Rectangular fit is another important criterion that can be used to capture the linear nature of windbreaks. According the ENVI Feature Extraction Module User’s Guide (ITT Visual Information Solutions 2008), rectangular fit can be used to eliminate circular and radically ‘jagged’ features from any class when deemed necessary by the user.

Scale is also an important factor when setting segmentation parameters. In remote sensing “a certain scale is always presumed by pixel resolution” but “objects of interest often have their own inherent scale” (Benz et al., 2004 pg. 241). For example, while a windbreak may not be visible on a LANDSAT 5 image with 30 meter spatial resolution, it would be visible on VHR imagery with a spatial resolution of 1 meter. Though related, scale and spatial resolution do have distinct differences. Resolution refers to the “average area dimension a pixel covers on the ground,” where as scale measures the “magnitude or the level of aggregation (and abstraction) on which certain phenomenon can be described” (Benz et al., 2004 pg. 245).

(30)

Figure 5. Differences between the main and first-order lines of an object.

Ma

in

L

in

e

Main Line

Firs t ord er b ranc h

First order b_ranch

First order br anch Fi rs t o rd er b ran ch First o rder b ranch First o rder b ranch

Ma

in

L

in

e

Main Line

Firs t ord er b ranc h

First order b_ranch

First order br anch Fi rs t o rd er b ran ch First o rder b ranch First o rder b ranch

Once desired segmentation parameters have been determined, the next step in gathering useful information from an image is to place segmented objects into classes. Classifying objects simply means labeling certain image objects using category class names (Baatz et al., 2004). Classes, in turn, are established by placing like objects into a user-defined set of categories (Baatz et al., 2004). Fuzzy classification strategies, commonly referred to as fuzzy logic, assign a measure of membership to each pixel or object and have become increasingly popular when classifying objects (Wang 1990; Jager and Benz 2000; Blaschke et al., 2004). Fuzzy logic can also be used to assign certain objects or shapes to an individual class. Membership values fall

(31)

between 0.0 and 1.0, with 0.0 indicating no membership exists with any given class and 1.0 means full membership (Baatz et al., 2004).

Jager and Benz (2000) suggest the fuzzy logic approach has great value when compared with ground truth data collected for accuracy measures because it takes into account mixed pixel issues by assigning a degree of membership to each relevant category. Certainly, fuzzy logic would be useful for identifying windbreaks as such objects will frequently have mixed pixels due to their narrow linear character that can result in frequent overlap of forest pixels with neighboring grasses or other landcover types. Also, many windbreaks are planted in distinct rows that permit soil and grasses to be visible within them in addition to gaps caused by the presence of dead trees.

Using a supervised classification strategy, multiple objects can be selected as training areas for each class to be established. Various shape, contextual, spectral, and other spatial attributes, such as length and area, can then be calculated for and assigned to training sites to determine what other objects in the image fit into the predetermined classes best (Baatz et al., 2004). Image processing software used in this research project incorporates both a supervised classification strategy and a rule-based classification strategy. The rule-based classification strategy can be used to isolate features of interest while eliminating irrelevant features and works by allowing the user to define certain criteria that objects must possess in order to be placed in a given class. It is essentially a singular classification that eliminates all but one particular set of objects.

This process is accomplished by first eliminating pixels that do not contain spectral values inherent to the features of interest. Second, shape and contextual based criteria can be used to further eliminate features not desired in the final output (ITT Visual Information

(32)

Solutions 2008). Recently, similar rule-based techniques have been used to identify urban features such as rooftops and buildings (Huang and Zhang 2008).

Once desired objects have been classified, they can be imported into a GIS as a vector or raster dataset for further analysis (Wiseman et al., 2007). According to Tian and Chen (2007), however, vectors exported from certain image processing platforms sometimes contain no spatial reference. This prevents a GIS from recognizing the actual size, spatial extent, or location of the data. To address this issue, Tian and Chen (2007) developed a custom tool that allowed them to set the spatial reference. Spatial reference definitions did not appear to be an issue with research done using more modern image processing software (Radoux and Defourny 2007; Huang and Zhang 2008; Rutherford and Rapoza 2008).

Object-based Classification Implementation

Traditional methods of classification using remotely-sensed images have become time consuming and inefficient given the development of automated classification techniques (Drăgut and Blaschke 2006). In fact, even traditional pixel-based classification techniques are being replaced with object-based classification due to its increased accuracy and versatility. Whiteside and Ahmad (2005) found that object-based image analysis produced 78% accuracy while pixel-based approaches yielded 69.1% in a land cover classification comparison in northern Australia. Koch et al., (2007) classified landuse and landcover at a study site within the interior Atlantic forest of Paraguay in an attempt to monitor Hantavirus dynamics. Their research showed that object-based classification produced 84% overall accuracy while the per-pixel approach yielded only 43% overall accuracy (Koch et al., 2007). Koch et al., (2007) also reported that the Kappa metric calculated from their results was significantly higher for the object-based classification results compared to that of the per-pixel results. Cohen’s Kappa is a statistic used to measure

(33)

how much better the classification worked compared to random chance assignment of features to classes (Jenson 2005). Higher percentage accuracy for object-based classifications was a common trend throughout all of the literature reviewed for this research. The topic of accuracy assessment is addressed in detail by Jäger and Benz (2000) and Jenson (2005 pg. 506-509).

Object-based classification can be used for a variety of image analysis objectives. Commonly, object-based classification is used when classifying land cover, however, it is certainly not limited to that function. Baatz et al., (2006) used object-based classification for high content screening of fluorescent cell images at a microbiological level. Geomorphologists such as Drăgut and Blaschke (2006) have used object-based classification to identify landform elements and created a reproducible methodology that could be applied to a number of study sites. Most recently, researchers have been using an object-based approach to classify urban areas which are inherently hard to classify due to the diverse range of spectral values that are typically present (Huang and Zhang 2008; De Roeck et al., 2009; Bernad et al., 2009). In fact, De Roeck et al., (2009) were able to produce a Kappa value of 0.84 when classifying urban fringe areas, meaning that their classification scheme worked 84% better than a random assignment of features to classes.

Relevant to this project is a study conducted by the Agriculture and Agri-Food Canada (AAFC) Prairie Farm Rehabilitation Administration (PFRA) Shelterbelt Centre in Indian Head, Saskatchewan. Similar to the KFS, the AAFC PFRA had no spatial data regarding the location of windbreaks, or their condition, within their area of responsibility. It was determined that traveling across the country to visit individual windbreaks would be too costly and time consuming. Therefore two main objectives were established for their research: (1) “Can high resolution imagery be used to generate an accurate inventory of shelterbelts across a vast

(34)

landscape?” and (2) “Can shelterbelts be identified by species from high resolution imagery?”(Wiseman et al., 2007 pg. 2). Wiseman et al., (2007) randomly selected 27 individual windbreaks for ground truth comparisons with windbreak features they extracted from aerial imagery. They did not classify windbreaks after segmenting their image but, instead, chose to export the vector-based polygon geometries into a GIS immediately after segmentation and isolate windbreaks from the rest of the polygons using standard query language (SQL) statements. Rather than classify the image, they used a set of shape and length SQL queries to pick out windbreaks from the rest of the segmented polygons. Using this technique, 26 of the 27 ground truth windbreaks were successfully identified. Wiseman et al., (2007) reported that shape attributes were most useful in identifying whole windbreaks, while spectral information produced better results for individual species classification. While their results seem accurate, it should be understood that Wiseman et al., (2007) did not employ a conventional classification strategy. Rather they used a technique more closely related to a process referred to as feature-based classification (Bernad et al., 2009).

In this scenario, features of interest are extracted from the rest of the polygons and placed into their own category or file. This means that there are no 1’s and 0’s, like in a binary classification, but just 1’s indicating the feature exists or “no data”. It should be understood that accuracy results from a single feature extraction such as this cannot be compared to results generated from a Kappa metric because they represent different measures. The Kappa metric is used to measure how much better a classification scheme worked compared to the random assignment of features to classes. Weisman et al., (2007) tried only to classify windbreaks that they knew to exist. The issue with their accuracy measure appears to be, that outside of the 27

(35)

windbreaks they visited in the field, they did not document whether any other windbreaks in their study area were correctly identified.

It has been noted that, in image classification, as more classes are developed the better the chances are of distinguishing between them (Jensen 2007). The idea is that more classes will eliminate confusion between features by reducing generalization that can occur when using too few classes. For example, if windbreaks were the only class selected for features that contain trees, then the all trees in the image would be placed into the windbreak class. However, if classes were developed for windbreaks, riparian zones, and larger forest stands, various algorithms can be used to place each “tree” feature in their correct respective classes based on other properties such as shape, size, and spectral reflectance.

Object-based classification has proven to be a useful technique in classifying land cover in a variety of scenarios. Often, object-based classification has shown its usefulness in classifying entire images and for single feature extraction. Using certain spatial and spectral criteria, it is the objective of this research to isolate windbreaks from all other land cover features and assess their condition using different spectral and textural properties. By taking advantage of the unique ability of object-based classification to classify features based on their shape, windbreak features should be easily distinguished from all other landcover features that share similar spectral properties. In addition, using this automated technique it should help decrease the amount of time required to inventory windbreaks by eliminating much of the field survey methods employed by Read (1958). While a variety of windbreak surveys have been conducted in the past, none performed in Kansas have ever attempted to extract windbreaks from aerial imagery at the county level.

(36)

CHAPTER 3 - Study Area

Ford County, Kansas was selected as the study area for this research project and will serve as the pilot county for the development of remote sensing and GIS methods to classify windbreaks (Figure 6). The methods tested in Ford County, if successful, will then be applied to a multi-county area in southwest Kansas, including the counties of Clark, Gray, Haskell, Hodgeman, Meade, and Seward. These counties are located in the Southwestern District of the KFS and selected for analysis in the Conserving, Renovating & Establishing Working Field Windbreaks (CREWFW) grant that funded this research. The CREWFW was accepted for funding via the State and Private Forestry FY 2008 Western Competitive Grant Application.

(37)

Ford County is home to more than 800 acres of windbreaks, some of which were planted during the PSFP (1930’s-1940’s). In addition, the Natural Resource Conservation Service (NRCS) provided an independently developed GIS-friendly spatial data file for windbreaks identified using a ‘heads-up digitizing’ process that will be used to assess the accuracy of the classification techniques applied in this project (R. Temat 2008, pers. comm.).

Ford County is located in southwest Kansas and spans an area of 2,850 km2. Landcover within the county is dominated by agricultural cropland, while mixed or short grass prairiesmake up the majority of native vegetation (Goodin et al., 2002 pg. 46). Figure 7 shows a map of Ford County landcover based on the 2001 National Landcover Dataset (NLCD). Land use related to the cattle industry also makes up a major part of the landscape with many areas being used as grazing land, livestock feed production, and large scale feedlots (Harrington 2001). A detailed breakdown of current landuse/landcover types in Ford County is shown in Table 2. Appendix A gives a description of the NLCD land cover types and associated cover codes.

Much of the county’s agricultural cropland remains bare soil after harvest, which usually occurs in mid- to late June. The lack of crop residue makes it extremely susceptible to wind erosion throughout much of the year. During the period between harvest and next planting, windbreaks sometimes provide the only protection against soil loss (Brandle et al., 1982).

Another important landcover feature in the study area is the lengthy riparian area bordering the Arkansas River. This forest corridor stretches approximately 65 km from the northwest to east-central border of the county. Eastern cottonwood (Populus deltoides) is the dominant tree species found in the riparian area, as well as in many of the older windbreaks (R. Atchison pers. comm. 2009).

(38)

Figure 7. Landcover/Landuse classes in Ford County.

Dodge City is the major urban center in Ford County and is home to 77% of the county’s residents (Harrington et al., 2003). Current population estimates from the U.S. Census Bureau (2009) shows this percentage holding steady, with Dodge City comprising 25,737 out of the county total population of 33,340. Table 3 shows historic population estimates from 1870 to the present (U.S. Census Bureau and National Historical GIS (NHGIS)). Drastic population spikes occurred between 1900-1930 as well as from 1960-2000. These population spikes are illustrated in Figure 8. Population booms in the 1930’s and 1960’s are due to the “Green Revolution”, an era in which advanced in agricultural technology made farming an appealing and profitable career choice (Evenson and Gollin 2003).

(39)

Table 2. Total area of each landuse/landcover class in Ford County (Source: 2001 NLCD)

COVER TYPE COVER

CODE

ACRES HECTARES % OF TOTAL

Open Water 11 19,021.2 7697.6 0.9%

Developed, Open Space 21 87,027.9 35,218.9 3.9%

Developed, Low Intensity 22 13,818 5,591.9 0.6%

Developed, Medium Intensity 23 3,377.2 1,366.7 0.2%

Developed, High Intensity 24 1,550.3 627.4 0.1%

Barren Land (Rock/Sand/Clay) 31 557.5 225.6 0%

Deciduous Forest 41 5,343.3 2,162.3 0.2% Evergreen Forest 42 12.7 5.1 0% Mixed Forest 43 19.1 7.7 0% Shrub/Scrub 52 69.1 27.9 0% Grassland/Herbaceous 71 564,014.8 228,248.6 25.4% Pasture/Hay 81 55,236.9 22,353.5 2.5% Cultivated Crops 82 1,455,024.7 588,827.6 65.6% Woody Wetlands 90 11,644.5 4,712.2 0.5%

Emergent Herbaceous Wetlands 95 492.4 199.3 0%

Total 2,217,209.4 897,272.8 100%

Table 3. Historic population change in Ford County, Kansas. Year Population Year Population

2009 33,340 2001 32,281 2008 33,293 2000 32,565 2007 33,077 1990 27,463 2006 32,751 1960 20,938 2005 32,876 1930 20,647 2004 32,654 1900 5,497 2003 32,558 1870 427 2002 32,164

(40)

Figure 8. Graph of population change in Ford County, 1870-2009.

Ford County KS Population Increace (1870-2009)

0 5000 10000 15000 20000 25000 30000 35000 40000 1870 1900 1930 1960 1990 2000 2009 Year P o pul a ti on

Ford County has a semi-arid climate characterized by high winds, low yearly precipitation (> 58 cm) and extreme weather events. Table 4 lists average annual precipitation values for the county dating back to 2000. Figure 9 is a graph of five-year mean precipitation totals plus the deviations from the long term average (47.9 cm) taken from the Dodge City weather station and date as far back as 1895. Periods of drought during the 1930’s are easily visible in Figure 9 as precipitation deviates well below the long term mean. In a tree ring chronology study under taken by Woodhouse and Overpeck (1998), it was determined that the severity of droughts during the 1930’s and 1950’s were equal to or slightly less than droughts experienced thousands of years ago. This study indicates that severe drought has consistently plagued the study area for hundreds of years, with some drought periods lasting for decades or longer (Woodhouse and Overpeck 1998).

(41)

Table 4. Annual precipitation for Ford County, 2000-2008.

Year Precipitation (in) Precipitation (cm)

2000 23.3 59.3 2001 17.5 44.4 2002 17.8 45.3 2003 19.1 48.4 2004 23.5 59.7 2005 25.2 64.1 2006 24.6 62.4 2007 27.2 69.1 2008 22.7 57.6

Figure 9. Five year precipitation averages for Ford County, 1895-2006.

5 Year Precipitation Average (1895-2008)

-15 -5 5 15 25 35 45 55 65 1899 1909 191 9 192 9 193 9 1949 1959 1969 197 9 198 9 199 9 2008 Year Pre c ipitati o n (cm)

Dev. From Avg. 5 Year Avg

Blizzards are common events during winter months, while drought often strikes during the summer months. Severe winds, tornados, and thunderstorms accompanied with hail are also very common in the spring (Flora 1948). The National Climactic Data Center (2009) reported 73

(42)

documented tornado touchdowns between 1950-2008. Mean monthly temperature variations range from -1 o to +27o C (Goodin et al., 1995; Harrington et al., 2003). Figure 10 is graph of historic annual mean temperatures ranging from 1875-2008. This data was collected from the Dodge City weather station and made available by the Kansas Board of Agriculture.

Figure 10. Five year mean temperatures for Ford County, 1895-2006.

5 Year Average Temperature for Ford Co. KS (1895-2008)

10.5 11 11.5 12 12.5 13 13.5 1899 190 4 190 9 191 4 191 9 192 4 192 9 1934 193 9 1944 194 9 1954 1959 1964 1969 1974 1979 198 4 1989 199 4 1999 200 4 200 8 Year T e m p er at u re ( C )

High winds accompanied by severe weather events (e.g., blizzards) are common to the area and have the potential to cause significant damage to shelterbelts. Winter ice storms in 2007 caused major widespread damage to windbreaks in the area due to the combined impact of accumulated ice and winds which broke limbs and, in some instances, brought down entire trees (R. Atchison. pers comm. 2009).

(43)

CHAPTER 4 - Data and Methods

Imagery

The image used in this study was acquired through the 2008 National Agricultural Imagery Program (NAIP). NAIP imagery has a spatial and radiometric resolution of 1 m and 8 bits, respectively (United States Department of Agriculture 2008). Each image is multispectral in nature, and contains spectral data from four bands of the electromagnetic spectrum: Blue (band 1), green (band 2), red (band 3), and near-infrared (NIR) (band 4). Documentation was not available with the NAIP image that described the specific wavelengths recorded in each band. The NAIP image used here was collected during the crop-growing months of 2008.

Individual multispectral NAIP images are processed into Compressed County Mosaics (CCMs), with a compression ratio of 1:15, and distributed to the public in the JPEG 2000 (.jp2) format (Adkins 2008). The 2008 NAIP CCM for Ford County was downloaded from the Kansas Geospatial Community Commons (http://www.kansasgis.org). The image is delivered to the public with a maximum of 10% cloud cover, georectified (North American Datum 1983 Universal Transverse Mercator Zone 14 North) and with atmospheric corrections already made (Mathews and Davis 2007). Because little cloud cover was observed in the Ford County image, it was determined that the few clouds that did exist would not introduce significant error into later windbreak classification efforts.

One great advantage of using the 2008 NAIP imagery is that it can be downloaded for free as compared to similar 1 m resolution imagery which can cost over $25.00 U.S. dollars per km2 (M. Kallas, 2008 pers. comm.). Because the area of Ford County is 2,850 km2, the total cost associated with purchasing imagery could easily be $71,250, making it unaffordable for this and other projects.

(44)

The remainder of this chapter outlines methods used to process the NAIP image, perform object-based classification, assess windbreak condition, and calculate the area of windbreaks found on highly erodible soils. Figure 11 diagrams a general workflow of the methods used in this research.

Figure 11. Workflow sketch of the methods used in this research.

Image Preprocessing

Most GIS software packages are compatible with the .jp2 compressed file format. However, the digital image processing software program used in this study, ENVI Zoom 4.5 Feature Extraction Module (ITT Visual Information Solutions, Boulder, CO), did not recognize

(45)

the .jp2 format. Because of this, the original .jp2 CCM for Ford County was converted to the TIFF format (based on a lossless compression technique) to make it universally compatible with the all image processing and GIS software packages used here.

Image resampling was performed next and the process of geometrically transforming an image and is commonly used to enhance the visual characteristics of an image for display purposes (Parker et al., 1983; Dodgson 1992). Image resampling can also be used to decrease the spatial resolution of an image by enlarging the dimensions of individual pixels through resampling techniques such as nearest neighbor, bilinear interpolation, and cubic convolution filtering. This process is sometimes referred to as spatial down-sampling (Parker et al., 1983; ITT Visual Information Solutions 2008).

The advantage of using fine spatial resolution imagery comes at the cost of enormous file sizes which can significantly slow computer-based processing methods. Dodgson (1992) suggests that image resampling should be done when, (1) collecting the imagery at different resolutions is not an option, and (2) when it is possible to recapture the imagery but resampling requires no additional monetary input and would result in faster processing. Here we cannot have NAIP personnel re-fly Ford County to capture imagery with less spatial resolution and faster processing times are obviously desired. The original .jp2 NAIP image was approximately 14.5 GB which exceed the 4 GB limit to convert to a .tiff file (Geospatial Data Abstraction Library 2009).

Based on Dodgson (1992) suggestions it was confirmed that resampling the original imagery was indeed the best option. Though information is lost in this process, it will not impact this research. For example, a data reduction step such as resampling would not be appropriate if individual tree species were to be classified. Here, however, the objective is to identify entire

(46)

windbreaks, so image resampling is actually beneficial by both reducing file sizes and by transforming windbreaks into areas of more homogeneous pixels rather than isolated pixels representing single trees with in a windbreak.

The Nearest Neighbor (NN) resampling technique was used because of its speed and simplicity (Parker et al., 1983, OverWatch Textron Systems 2008). The Feature Analyst Reference Manual (OverWatch Textron Systems 2008) also suggests that the NN resampling technique be used when linear features, like windbreaks, are of particular interest.

Resampling factors of 2, 4, 6, 8, and 10 were all performed to determine an optimal solution to the .tiff conversion issue. Resampling the image at factors of 2 and 4 did not decrease the original .jp2 file size sufficiently to allow for file conversion. In addition, when resampling factors of 8 and 10 were used windbreaks in the imagery appeared to lose most, if not all, spatial integrity. Pixels comprising small windbreaks were averaged in with surrounding grasslands and all digital information regarding them was lost. Furthermore, when resampling factors of 8 and 10 were used, several large field windbreaks were averaged into a single row of pixels which would have prevented an accurate assessment of their condition in later classification steps.

Resampling the image by a factor of 6 reduced the .jp2 file size enough to permit for .tiff conversion while maintaining much of the spatial integrity of field and farmstead windbreaks. Figure 12 illustrates a sample of the original and resampled image that was used for segmentation and classification.

(47)

Figure 12. Comparison of the Ford County 2008 NAIP image at 1 and 6 meter pixel size.

Image Segmentation and Merging

Several segmentation methods are available that can be used to isolate homogeneous pixels into proper objects. In a preliminary segmentation test for this project conducted as part of a class assignment, a bottom-up region merging approach was used to segment imagery within a test site in Ford County. Using this approach, and a minimum distance classifier, an overall accuracy of 53.2% was achieved for a landuse/landcover classification. For this research, an edge-based segmentation method and Support Vector Machine (SVM) classifier was applied to determine whether improved accuracy was possible.

Segmenting imagery can be a time-consuming activity depending on the processing power of the computer hardware being used to operate the segmentation software. Because of this, an area of just over 240 km2 was extracted from the resampled imagery to develop segmentation and classification parameters to be applied later to the entire CCM of Ford County (Figure 13). Developing parameter settings using a smaller image file allows for faster

(48)

on the much larger CCM file. The test site was selected because it contained most of the windbreak ground truth data collected for accuracy assessment. In addition, this particular area was noted as having a particularly high number of windbreaks (> 30) compared to other areas of the county.

Figure 13. Location of Ford County test site.

Segmenting an image relies heavily on a trial-and-error approach where the user evaluates the performance of several different scale parameter settings in order to determine which performs best. The ENVI Feature Extraction Module User’s Guide (ITT Visual Information Solutions 2008) suggests users should implement the highest scale level that properly defines the boundaries of features of interest. The image processing software used in this research provided a viewing portal in which different segmentation parameters could be set

(49)

and the results then viewed “on the fly” prior to segmenting the entire image to determine if the segments are satisfactory. The viewing portal is mobile, meaning users can move the portal to any region of the image and view a sample of what segments would result.

Scale level is an area measure that determines the ‘size’ of objects to be created. Scale parameter values range from 0.0-100.0 with segments decreasing in size as they move closer to 0.0. After viewing several scale parameters in the viewing portal, it was determined that a scale parameter of 78 adequately delineated large homogeneous regions within windbreaks throughout the image (Figure 14). Higher scale parameters appeared to create larger objects that encompassed more landuse/landcover types than just windbreaks. Lower parameters values began to create borders around individual trees within windbreaks.

Figure 14. Screenshot of the ENVI Zoom 4.5 Feature Extraction Module and viewing portal (outlined in red).

(50)

Using a scale level of 78, an edge-based segmentation algorithm was used to segment the entire Ford County image. The edge-based segmentation algorithm works faster than other approaches, such as bottom-up region merging, because it requires only one parameter input (scale level). Though the scale level of 78 adequately defined segments between the windbreak and surrounding landcover, several small segments within the windbreak remain. To correct for this issue, a technique called segment merging was used to group small segments of similar spectral, spatial, and textural values to create even larger segments containing similar values.

Segment merging is based on a Full Lambda-Schedule algorithm created by Robinson et al., (2002) (Equation 1). Merging of objects occurs when the algorithm identifies a set of neighboring objects with similar spatial and spectral properties. The merge which groups shapes into even larger objects uses the same Equation 1 as the segmentation process and has a range of threshold lambda values between 0.0 - 100.0. After experimenting with a wide variety of lambda values it was determined that a value of 50 properly merged segments within windbreaks while keeping them isolated from neighboring landuse/landcover types.

Equation 1. Full Lambda-Schedule algorithm for segment merging.

(

)

(

)

2 , , j i j i j i j i j i O O length u u O O O O t ∂ − ∗ + ∗ = where,

Oi = region i of the image, Oj = region j of the image, ui = average value in region i, uj = average value in region j,

||ui – uj|| = Euclidean distance between spectral values of regions i and i, and

Identifying and assessing windbreaks in Ford County, Kansas using object-based image analysis

Abstract

Table of Contents

List of Figures

List of Tables

List of Equations

Acknowledgements

Dedication

CHAPTER 1 - Introduction

Purpose and Objectives

Applied and Academic Contribution

CHAPTER 2 - Literature Review

Windbreaks

Previous Windbreak Assessments

Remote Sensing

Object-based Classification Concepts

Ma

in

L

in

e

Main Line

Ma

in

L

in

e

Main Line

Object-based Classification Implementation

CHAPTER 3 - Study Area

CHAPTER 4 - Data and Methods

Imagery

Image Preprocessing

Image Segmentation and Merging

(

)

(

)